update

Signed-off-by: shen-shanshan <[email protected]>

Author: shen-shanshan

vllm_ascend/worker/worker.py

@@ -17,6 +17,7 @@
 # Adapted from vllm-project/vllm/vllm/worker/worker.py
 #
 
+import gc
 import os
 from typing import Dict, List, Optional, Set, Tuple, Type, Union
 
@@ -304,7 +305,9 @@ def determine_num_available_blocks(self) -> Tuple[int, int]:
                              cache_block_size)
         num_npu_blocks = max(num_npu_blocks, 0)
         num_cpu_blocks = max(num_cpu_blocks, 0)
-
+        gc.collect()
+        # TODO: don`t need impl this func after empty_cache in
+        # Worker.determine_num_available_blocks() unified`
         NPUPlatform.empty_cache()
         return num_npu_blocks, num_cpu_blocks